147 results about "Restrict boltzmann machine" patented technology

A Deep Belief Network (DBN) feature extraction-based analogue circuit fault diagnosis method comprises the following steps: a time-domain response signal of a tested analogue circuit is acquired, where the acquired time-domain response signal is an output voltage signal of the tested analogue circuit; DBN-based feature extraction is performed on the acquired voltage signal, wherein learning rates of restricted Boltzmann machines in a DBN are optimized and acquired by virtue of a quantum-behaved particle swarm optimization (QPSO); a support vector machine (SVM)-based fault diagnosis model is constructed, wherein a penalty factor and a width factor of an SVM are optimized and acquired by virtue of the QPSO; and feature data of test data are input into the SVM-based fault diagnosis model, and a fault diagnosis result is output, where the feature data of the test data is generated by performing the DBN-based feature extraction on the test data.

In an approach to incident prediction and response, one or more computer processors receive one or more alerts corresponding to an incident. The one or more computer processors aggregate the one or more alerts with additional data corresponding to the incident. The one or more computer processors feed the aggregated data into a prediction model, where training of the prediction model uses associated independent stacked Restricted Boltzmann Machines utilizing one or more supervised methods and one or more unsupervised methods. The one or more computer processors determine, based, at least in part, on one or more calculations by the prediction model, at least one probability of the incident. The one or more computer processors determine whether the at least one probability exceeds a pre-defined threshold. In response to determining the at least one probability exceeds a pre-defined threshold, the one or more computer processors send at least one notification.

The invention proposes a rolling bearing fault prediction method based on a continuous deep belief network. The method comprises the following steps: first, extracting the time-domain characteristic quantity of vibration signals of a rolling bearing; then, fusing the extracted time-domain characteristic information using a locally linear embedding method, so as to define a new comprehensive monitoring index for better quantitative evaluation of performance degradation of the bearing; training a continuous restricted Boltzmann machine step by step to construct a continuous deep belief network prediction model; and using a genetic algorithm to optimize the structure of the continuous deep belief network so as to further improve the prediction precision. The prediction method is reliable in result, has good real-time performance, is simple and feasible, and is suitable for rolling bearing fault prediction.

The invention discloses a hyperspectral remote sensing data classification method based on deep learning, and belongs to the technical field of hyperspectral data classification. The invention aims to solve a problem of low classification precision of a method for classifying hyperspectral remote sensing data with nonlinear characteristics. The hyperspectral remote sensing data classification method comprises the following steps: firstly, processing hyperspectral original data to obtain the spectral feature vector and the spatial feature information of the hyperspectral original data; then, integrating the spectral feature vector with the spatial feature information; confirming labeled samples by hyperspectral integrated data, selecting a training sample and a test sample from the labeled samples; Pre-training a multi-layer restricted Boltzmann machine which forms a deep network by the training sample; carrying out supervised learning to the network formed by the multi-layer restricted Boltzmann machine through the training sample; and inputting the test sample into the trimmed network formed by the multi-layer restricted Boltzmann machine to realize hyperspectral remote sensing data classification. The invention is used for the hyperspectral remote sensing data classification.

The invention discloses an advanced persistent threat (APT) attack detection method based on deep belief network-support vector data description. A deep belief network (DBN) is used for feature dimension-reduction and excellent feature vector extraction; and support vector data description (SVDD) is used for the data classification and detection. At a DBN training state, the feature dimension-reduction is performed by using the DBN model after obtaining a standard data set; a low-level restricted Boltzmann machine (RBM) receives simple representation transmitted from the low-level RBM by usingthe high-level RBM so as to learn more abstract and complex representation after performing the initial dimension-reduction, and back propagation of a back propagation (BP) neural network is used forrepeatedly adjusting a weight value until the data with excellent feature is extracted. The data processed by the DBN is divided into a training set and a testing set, and the data set is provided for the SVDD to perform training and identification detection, thereby obtaining the detection result. The attack detection method disclosed by the invention is suitable for the unsupervised attack datadetection with large data size and high-dimension feature, is fit for the APT attack detection and can obtain an excellent detection result.

The invention provides a video heat prediction method based on deep belief networks. The method comprises following steps: selecting input variables according to video features; normalizing and quantifying influence factors, thus preprocessing training data; determining a single-layer restricted Boltzmann machine reconstitution dimension according to the selected input variable and feature reconstruction errors; forming the deep belief networks by multi-layer stacked restricted Boltzmann machines and a BP (Back Propagation) neural network; adjusting the deep belief networks through a global learning algorithm, thus obtaining an optimum video prediction model; placing to-be-tested video test data in the optimum video prediction model for heat prediction analysis and watching quantity prediction analysis. The invention also provides a video heat prediction system based on the deep belief networks. According to the method and the system of the invention, an online video prediction model based on the deep belief networks is provided; the deep neural network is applied in the online video prediction field; and the prediction accuracy and reliability can be improved.

The embodiment of the invention, which relates to the technical field of the computer, provides a voice recognition method and apparatus, so that a problem of a low voice classification distinguishing degree during the voice classification process according to the existing voice recognition technology can be solved. The method comprises: at least two voice features of a to-be-recognized voice are extracted; on the basis of a multi-layer restricted Boltzmann machine (RBM), each of the at least two voice features is trained to obtain a depth voice feature corresponding to each voice feature; feature fusion is carried out on the depth voice feature corresponding to each voice feature, thereby obtaining a depth voice feature of the to-be-recognized voice; and the depth voice feature of the to-be-recognized voice is inputted into a classifier and is classified, so that a voice type of the to-be-recognized voice is obtained. The method and apparatus are applied to voice recognition.

The invention relates to a method for re-identifying persons on the basis of deep learning encoding models. The method includes steps of firstly, encoding initial SIFT features in bottom-up modes by the aid of unsupervised RBM (restricted Boltzmann machine) networks to obtain visual dictionaries; secondly, carrying out supervised fine adjustment on integral network parameters in top-down modes; thirdly, carrying out supervised fine adjustment on the initial visual dictionaries by the aid of error back propagation and acquiring new image expression modes, namely, image deep learning representation vectors, of video images; fourthly, training linear SVM (support vector machine) classifiers by the aid of the image deep learning representation vectors so as to classify and identify pedestrians. The method has the advantages that the problems of poor effects and low robustness due to poor surveillance video quality and viewing angle and illumination difference of the traditional technologies for extracting features and the problem of high computational complexity of the traditional classifiers can be effectively solved by the aid of the method; the person target detection accuracy and the feature expression performance can be effectively improved, and the pedestrians in surveillance video can be efficiently identified.

The invention provides a natural gas pipeline tiny leakage detection method based on a sound signals. By taking the sound signal fluctuating most intensely during pipeline tiny leakage as a main signal, and temperature, pressure and flow rate signals as secondary signals, a deep convolution neural net based on multiparameter is established, and the probability of tiny leakage of a natural gas pipeline is determined through studying data of the main signal and the secondary signals; in addition, a deep restricted boltzmann machine model is designed specifically for the isomerism of multiparameter data, the multiparameter data is subjected to pre-training, an optimal weight value matrix and an optimal offset value are obtained and act as a weight initial value and an offset value initial value of a deep convolution neural net model. With the adoption of the method, an accuracy rate of pipeline tiny leakage accident judgment is increased greatly, the probability of a large safety accident is reduced, and the rescue efficiency of a natural gas pipeline accident is improved.

The invention discloses an isolated digit speech recognition classification system and method combining a principal component analysis (PCA) with a restricted Boltzmann machine (RBM). First of all, a Mel frequency cepstrum coefficient (MFCC) is employed for combination with a one-order difference MFCC, and a voice dynamic characteristic of an isolated digit is preliminarily drawn off; then, linear dimension reduction processing is carried out on an MFCC combination characteristic by use of the PCA, and dimensions of a newly obtained characteristic are unified; accordingly, nonlinear dimension reduction processing is performed on the obtained new characteristic by use of the RBM; and finally, finishing recognition classification on a digit voice characteristic after nonlinear dimension reduction by use of a Softmax classifier. According to the invention, PCA linear dimension reduction, unification of the dimensions of the characteristic and RBM nonlinear dimension reduction are combined together, such that the characteristic representation and classification capabilities of a model are greatly improved, the isolated digit voice recognition correct rate is improved, and an efficient solution is provided for high-accuracy recognition of isolated digit voice.

The invention relates to a risk identification method and system based on transfer deep learning. The risk identification method comprises the steps of generating vectors for all possible features through prescribed preprocessing, enabling the vector set to act as visible layer input of a first RBM ((Restricted Boltzmann Machine) so as to build an RBM layer; performing transfer learning by using known fraud samples, and carrying out transfer weighted BP tuning on the RBM layer built in RBM building step; and judging whether the RBM meets prescribed conditions or not after BP tuning, if the RBMmeets the prescribed conditions, not requiring to increase an RBM layer and continuing the following step, and if the RBM does not meet the prescribed conditions, repeating the steps of RBM buildingand transfer weighted BP tuning. A judgment model can be built more accurately and emerging fraud means can be better dealt with according to the invention.

Embodiments of the present invention provide an enhanced Restricted Boltzmann Machine (RBM) system with a novel regularization term to generate features automatically that are suitable for predicting remaining useful life (RUL) of engineered systems such as machines, tools, apparatus, or parts. The system improves the trendability of the output features, which may better represent the degradation pattern of such systems. The disclosed system has been demonstrated to improve trendability and RUL prediction accuracy, offering improved predictive power earlier in the life cycle of the machine, tool, or part. During operation, the system implements an RBM including a loss function. The system then extracts a set of features from a degradation measurement via the RBM. The system fits a rate-of-change slope for a respective feature and adds a regularization term to the loss function based on the fitted slope. The system then selects a subset of the enhanced features based on a measure of monotonic trending and aggregates the subset into a health value. The system then predicts a RUL as a weighted average of features best matching a historical degradation pattern in the health value.

The invention discloses a deep neural network training method based on multiple pre-training. According to the method, the pre-training process of a deep neural network is completed by performing pre-training with a restricted Boltzmann machine and a denoising auto-encoder acting as models in turn. With application of the method, the result of pre-training, i.e. the initial value of the deep neural network parameters, is enabled to have the characteristics of different models so that a better initial model is provided for later training. The method only restricts the value of the parameters of the beginning of different pre-training phases so that the method can be extended to other layer-by-layer pre-training models, such as the auto-encoder and a contraction auto-encoder, etc.

The invention discloses a deep neural network-based SAR (Synthetic Aperture Radar) texture image classification method, and aims to mainly solve the problem of low accuracy of SAR texture image classification with a larger number of samples and more characteristic dimensions in the prior art. The method is implemented by the following steps: (1) extracting low-level characteristics of an SAR image; (2) training the low-level characteristics of the SAR image to obtain advanced characteristics of the image by virtue of a first layer of RBF (Radial Basis Function) neural network of a deep neural network; (3) training the advanced characteristics to obtain more advanced characteristics of the image by virtue of a second layer of RBM (Restricted Boltzmann Machine) neural network of the deep neural network; (4) training the more advanced characteristics to obtain image texture classification characteristics by virtue of a third layer of RBF neural network of the deep neural network; (5) comparing texture classification characteristics of an image test sample with a test sample tag, and regulating parameters of each layer of the deep neural network to obtain the optimal test classification accuracy. The method is high in classification accuracy, and can be used for target identification or target tracking.

The invention discloses a fault classification model and method based on a sparse Gaussian Bernoulli restricted Boltzmann machine and a recurrent neural network. The fault classification model is composed of the sparse Gaussian Bernoulli restricted Boltzmann machine, a long and short time memory recurrent neural network, a perceptron and Softmax output layer; the sparse Gaussian Bernoulli restricted Boltzmann machine can learn nonlinear characteristics of data from unlabeled data, the recurrent neural network can process sequence data well, and solves the problem of gradient disappearance or gradient explosion occurring in a network training process by using a long and short time memory unit, and the perceptron and the Softmax output layer enhance the supervised classification ability of the network. The model of the invention has excellent characteristic extraction and perception ability of nonlinear data and dynamic data, and can well solve the problem of low fault classification accuracy caused by the nonlinearity of process data and the dynamics of fault data.

The invention discloses a power load prediction method based on a deep belief network. The power load prediction method comprises the following steps: aggregating historical data of a power load by adopting a sparse self-encoding neural network; constructing a composite optimized deep belief network prediction model based on a restricted Boltzmann machine, wherein the deep belief network prediction model sequentially comprises a Gaussian-Bernoulli restricted Boltzmann machine, a Bernoulli-Bernoulli restricted Boltzmann machine and a linear regression output layer from input to output, carryingout pre-training by adopting an unsupervised training method, and then carrying out parameter fine tuning by adopting a BP algorithm added with impulse terms; and inputting the aggregated historicaldata into the deep belief network prediction model for prediction. The invention further discloses a power load prediction system based on the deep belief network. According to the method, the regularity of historical load data can be better mined, so that the prediction efficiency is improved, meanwhile, the influence of different factors can be fully considered, and the prediction precision is improved.

The invention discloses a sequence deeply convinced network-based pedestrian identifying method. The method comprises the following steps of preprocessing a training image in a pedestrian database to obtain a training sample image, extracting an HOG (Histograms of Oriented Gradients) feature from the obtained training sample image, building and training a sequence restricted Boltzmann machine-based sequence deeply convinced network, using the sequence deeply convinced network to further extract features from the obtained HOG feature to form a feature vector of the training sample, inputting the obtained feature data into a support vector machine classifier, and finishing training; preprocessing a to-be-tested pedestrian image to obtain a test sample; using an HOG and the sequence deeply convinced network to extract pedestrian features from the test sample to form a feature vector of the test sample; inputting the feature vector of the test sample into the support vector machine classifier, and identifying whether the test image is a pedestrian or not. According to the method, better classification performance can be obtained, the accuracy of pedestrian identification is improved, and the robustness of a pedestrian identifying algorithm is enhanced.

The invention discloses an accelerating degradation model for a photovoltaic component constructed on the basis of a deep learning method and a forecasting method for the service life of the photovoltaic component. The method comprises the following steps: constructing a DNN (Deep Neural Network) through an RBM (Restricted Boltzmann Machine); taking different acceleration stress conditions (Ti, Hi and Rai) and a corresponding false failure service life distribution quantile function Qi(p) as input vectors; training RBM and DNN by utilizing a CD quick learning algorithm; searching for an optimal model parameter set theta*; and constructing the accelerating degradation model for the photovoltaic component, thereby forecasting the expected service life of the photovoltaic component under a normal stress condition.

The invention discloses an identity recognition method based on a gait image. The method includes: a training step: performing pedestrian detection and image preprocessing on a gait image sequence, assigning tag values to corresponding gait images, performing training by employing a characteristic learning network formed by a convolutional restricted Boltzmann machine and a full connection layer,and generating a characteristic learning network model with identity recognition and a characteristic central value model; and a recognition step: performing pedestrian detection and image preprocessing on a to-be-recognized gait image, calculating the gait periodicity through a normalized autocorrelation function to obtain a gait sequence of one period, and recognizing the identity of a pedestrian through a deep learning network and a voting algorithm. According to the method, the periodic gait image sequence is regarded as input, and complete gait information is reserved; characteristic learning is performed by employing the deep learning network, and more gait characteristics with a distinguishing degree are obtained to improve the recognition rate; and the recognition accuracy and robustness can be improved through combination usage of the deep learning network and the voting algorithm.

The invention discloses a method for predicting the residual life of a proton exchange membrane fuel cell. Firstly, obtaining output voltages of the proton exchange fuel cell at a plurality of continuous moments; constructing a training sample set, constructing a proton exchange membrane fuel cell residual life prediction model including one input layer, three hidden layers and one output layer; adopting an input layer and three hidden layers to construct a deep belief network model constructed by stacking three Gaussian restricted Boltzmann machines; and constructing an extreme learning machine model by adopting the last two hidden layers and the output layer, training the residual life prediction model by adopting the training sample set, acquiring the output voltage of the proton exchange membrane fuel cell at the latest moment, and predicting the residual service life of the fuel cell through the residual life prediction model. The method can effectively improve the accuracy and stability of the prediction result of the residual life of the proton exchange membrane fuel cell.

The utility model relates to a DBN (Deep Belief Network) based ADHD (Attention Deficit Hyperactivity Disorder) discriminatory analysis method. The ADHD discriminatory analysis method comprises the following steps: step 1, pre-processing; step 2, characteristic extracting and classifying: depending on the DBN that is formed by stacking RBMs (Restricted Boltzmann Machines), classified and reversely adjusted in a layer-by-layer manner through softmax finally. The targets of the RBMs in layer-by-layer training are to maximize the likelihood function of the probability function, to introduce in the comparison divergence, and to update the weight function, so that the hidden layer becomes the approximate representation of the visible layer, the hidden layer of the first layer serves as the visible layer of the second layer, by parity of reasoning, the RBM layers of the DBN are obtained, and the last hidden layer is adopted as the input of the softmax to obtain the corresponding output, namely, the classification. The adopted DBN is a probability generative model, is formed by stacking the multiple RBMs with the hidden layers and the visible layers, simulates the layer-by-layer abstract characteristic process when the human brain processes signals, and abstracts the equivalent characteristic expression of the original signals to apply in the field of ADHD classification.

The invention relates to an injurious insect image automatic recognition method based on a deep restricted Boltzmann machine. The method comprises a training process and a testing process. According to the training process, data of a training image set are preprocessed, preprocessed training images are grouped to construct a training image cube, features of each training image group are extracted through a restricted Boltzmann machine algorithm, and trained training image set feature data are obtained through feedback adjustment. According to the testing process, a test image to be recognized is input and is preprocessed, features of the test image are extracted through the restricted Boltzmann machine algorithm, and feature data of the test image with small errors are obtained through feedback adjustment; the varieties of injurious insects are recognized, and a preventive method is given out. The invention discloses an injurious insect image automatic recognition system based on the deep restricted Boltzmann machine. The injurious insect recognition rate and procedure robustness are improved, and the actual application value of injurious insect recognition in agricultural production is improved.

The invention provides an industrial process data modeling method for online self-adaptive fine-tuning deep learning. The method comprises the following steps: 1, acquiring process variables related to quality in an industrial process according to a sampling time sequence, labeling the process variables by utilizing the quality variables, constructing a labeled sample set, preprocessing data according to data characteristics, and dividing the labeled sample set into a training sample set and a query sample set; 2, training L restricted Boltzmann machines offline layer by layer in an unsupervised pre-training mode by using the training sample set, and stacking the L pre-trained restricted Boltzmann machines in sequence to form a stacked network. According to the industrial process data modeling method for online adaptive fine-tuning deep learning, the samples are selected based on similarity measurement to perform online updating on the offline model, the exclusive online model is established for the query samples, the accurate quality prediction of the query samples is completed, and the performance and precision of online prediction are improved.

The invention discloses a fault classification model and method based on a stacked sparse Gaussian Bernoulli restricted Boltzmann machine and reinforcement learning. The fault classification model iscomposed of a reinforcement learning Q-learning method and a sparse depth belief network model; an SDBN network is trained layer by layer in a non-supervision manner, then gradient descent training isperformed on the entire network in combination with the reinforcement learning Q-learning method, the network weight is adjusted, the correlation between adjacent sampling points between samples andthe dynamic characteristics of process data on a timing sequence are fully considered, and the ability of the model to extract the features of nonlinear and dynamic data in the process is further improved, so that the accuracy of fault classification is improved. The fault classification model disclosed by the invention can effectively solve the problem of low fault classification accuracy rate caused by the nonlinearity of the process data and the dynamics of the fault data.

The invention relates to a logging interpretation method related to lithology identification, in particular to a formation lithology identification method based on a deep belief network. The logging lithology identification method based on the deep belief network is mainly completed through a computer, and equipment needed for achieving the method comprises a logging instrument, a data communication interface and a computer. The method comprises the following steps: identifying lithology around a wellhead by utilizing logging data; preprocessing the logging data: performing normalization processing; digitalizing the lithology classification; calculating the correlation degree between the logging curve and the lithology; presetting the structure of the deep belief network; determining the number of restricted Boltzmann machines; determining a lithology classification boundary; training the deep belief network used for identifying formation lithology; and inputting the well logging dataof the well to be interpreted into the network, and carrying out lithology identification work. The identification method provided by the invention is simple in prediction, high in identification accuracy, good in effect, practical and reliable for regions lacking stratum element logging and imaging logging data.

A method is disclosed herein that includes an act of causing a processor to access a deep-structured, layered or hierarchical model, called deep convex network, retained in a computer-readable medium, wherein the deep-structured model comprises a plurality of layers with weights assigned thereto. This layered model can produce the output serving as the scores to combine with transition probabilities between states in a hidden Markov model and language model scores to form a full speech recognizer. The method makes joint use of nonlinear random projections and RBM weights, and it stacks a lower module's output with the raw data to establish its immediately higher module. Batch-based, convex optimization is performed to learn a portion of the deep convex network's weights, rendering it appropriate for parallel computation to accomplish the training. The method can further include the act of jointly substantially optimizing the weights, the transition probabilities, and the language model scores of the deep-structured model using the optimization criterion based on a sequence rather than a set of unrelated frames.